Research On Fall Detection Based On Acceleration Sensor

In recent years,the aging of China's population has become increasingly severe.In addition to the one-child policy,the proportion of empty-nest old people has increased year by year.These elderly people cannot be monitored for 24 hours.Studies have shown that falls are a major factor leading to the disability and death of the elderly.If an empty nester can't stand up alone after falling,it is often difficult to send out a distress message.When a family member discovers it,he usually misses the golden rescue time.Therefore,it is particularly important to develop a fall detection system to better protect empty nesters.Existing fall detection methods have poor detection performance for "new category" samples that do not exist in the training set.In order to better solve this problem,we have studied and developed a set of fall detection system based on triaxial acceleration sensors.The main content of this thesis is as follows:Firstly,a fall detection system is designed and built.The system consists of an external acceleration sensor,smart phone and server.The external sensor is worn at the waist and collects user motion information in real time,and transmits the data to the smartphone via Bluetooth.The smart phone uses the Internet to transmit data to the server to perform a fall detection on the server.Secondly,in order to construct a fall detection classification model,it is necessary to train the model of the classifier through a self-built data set.A total of 6 volunteers participated in the data collection.The volunteers simulated 11 kinds of activites of daily living and 9 kinds of activites of fall.In addition,in order to verify the generalization ability of the detection method,this paper also uses three public data sets to verify the algorithm.Thirdly,designed a fall detection method based on hidden Markov model.Fall detection is actually a two-category problem.This method needs to train two Hidden Markov Models for each type of data.When testing,the sample's probability of being generated in the two models are compared.The category with higher probability is the behavior of the sample.An oversampling method that preserves covariance was used to solve the class imbalance problem for time series.In this paper,two feature extraction equations are proposed for this method,and the continuous time series is discretized,and it is proved that the mean value dimension reduction does not affect the classification effect of the detection method.Fourthly,designed a fall detection method based on support vector machine.Based on the difference between the activites of daily living and the activites of fall,seven feature extraction equations were proposed,which improved the classification performance of the"new category" sample by the fall detection method.In order to select the feature combinations that have the best classification results,this paper carries out the sequential backward feature selection experiment.At the same time,we also performed support vector machine classification experiments on unextracted feature data,and proved that the feature extraction significantly improved the experimental results.The experimental results show that the hidden Markov model does not have good classification effect for datasets with more complex motion types,but it can achieve better classification effects for datasets with less classes.The result of the fall detection experiment based on the support vector machine is better than the fall detection experiment result based on the hidden Markov model.On the five datasets,the sensitivity and specificity of the support vector machine classification results were all higher than 96%.The fall detection method based on the support vector machine still has a good classification performance for the action categories that do not exist in the training set.The sensitivity and specificity is higher than 94%.